Source: OpenAI Prompt Caching Guide
OpenAI API guide on automatic prompt caching: prefix routing, prompt_cache_key, retention tiers, TTL pricing on GPT-5.6+, and usage fields.
Contents
- What: official API documentation page
- Author/publisher: OpenAI
- Date: living document; read 2026-11-14
- Original: Prompt caching
Summary
OpenAI caches prompt prefixes automatically for supported models; no request change is needed. The guide explains how requests are routed to machines that may hold a matching prefix, how to influence routing with prompt_cache_key, the retention policies for older models, and the newer explicit-breakpoint and TTL model introduced with GPT-5.6.
Key claims
- Caching is on by default and covers the full rendered prefix: instructions, tool definitions, conversation history including images, documents and audio.
- GPT-5.6 and later: minimum cacheable prefix is 1,024 visible tokens, exact boundaries reported; earlier models have model-dependent minimums and report
cached_tokensin multiples of 128. - GPT-5.6+:
prompt_cache_options.modeisimplicit(breakpoint at end of latest message) orexplicit(developer-placed breakpoints); up to four cache writes per request. - Caches live on individual machines. Routing uses a hash of the initial prefix tokens plus
prompt_cache_key; more than about 15 requests per minute for one prefix+key can overflow to other machines and miss. - Older models:
prompt_cache_retention=in_memory(typically 5–10 min idle, up to 1 h) or24h; reads billed at a discounted cached-input rate with no write surcharge. - Re-checked 2026-11-14: in
implicitmode the automatic breakpoint uses one of the four write slots; inexplicitmode a request with no breakpoints is not cached. The ~15 RPM per-key guidance applies to pre-GPT-5.6 models; on GPT-5.6+ routing is automatic andprompt_cache_keyis optional, used mainly for cache accounting. The 0.05× read rate applies to GPT-6.1 Sol and GPT-6.1 Terra. - GPT-5.6+:
ttlof30m(default, write 1.25× input) or12h(write 2.5× input, Responses API only); reads 0.1× input (0.05× on some GPT-6.1 models). The guide advertises discounts of up to 95%. - Usage:
usage.input_tokens_details.cached_tokens; GPT-5.6+ also reportscache_write_tokenssplit by TTL. - Changing model, tools,
parallel_tool_calls, output format, reasoning effort, verbosity or compaction settings can break the prefix. - The cache stores key/value tensors, not prompt text; compatible with Zero Data Retention; never shared across organizations; no manual clear. Cached tokens still count toward TPM rate limits.